Small Vectors, Big Effects: A Mechanistic Study of RL-Induced Reasoning via Steering Vectors

• URL: http://arxiv.org/abs/2509.06608v3
• Date: Wed, 01 Oct 2025 08:37:07 GMT
• Title: Small Vectors, Big Effects: A Mechanistic Study of RL-Induced Reasoning via Steering Vectors
• Authors: Viacheslav Sinii, Nikita Balagansky, Gleb Gerasimov, Daniil Laptev, Yaroslav Aksenov, Vadim Kurochkin, Alexey Gorbatovski, Boris Shaposhnikov, Daniil Gavrilov,
• Abstract summary: We study lightweight steering vectors inserted into the base model's residual stream and trained with a reinforcement-learning objective.<n>We find that (i) the last-layer steering vector acts like a token-substitution bias concentrated on the first generated token, consistently boosting tokens such as "To" and "Step"<n>We also show that steering vectors (i) transfer to other models, (ii) combine across layers when trained in isolation, and (iii) concentrate magnitude on meaningful prompt segments under adaptive token-wise scaling.
• Score: 12.331740215947677
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: The mechanisms by which reasoning training reshapes LLMs' internal computations remain unclear. We study lightweight steering vectors inserted into the base model's residual stream and trained with a reinforcement-learning objective. These vectors match full fine-tuning performance while preserving the interpretability of small, additive interventions. Using logit-lens readouts and path-patching analyses on two models, we find that (i) the last-layer steering vector acts like a token-substitution bias concentrated on the first generated token, consistently boosting tokens such as "To" and "Step"; (ii) the penultimate-layer vector leaves attention patterns largely intact and instead operates through the MLP and unembedding, preferentially up-weighting process words and structure symbols; and (iii) middle layers de-emphasize non-English tokens. Next, we show that a SAE isolates features associated with correct generations. We also show that steering vectors (i) transfer to other models, (ii) combine across layers when trained in isolation, and (iii) concentrate magnitude on meaningful prompt segments under adaptive token-wise scaling. Taken together, these results deepen understanding of how trained steering vectors shape computation and should inform future work in activation engineering and the study of reasoning models.

Related papers

• Attention Illuminates LLM Reasoning: The Preplan-and-Anchor Rhythm Enables Fine-Grained Policy Optimization [56.083511902353365]
  Reinforcement learning (RL) typically applies uniform credit across an entire generation of Large language models.<n>This work positions attention as a privileged substrate that renders the internal logic of LLMs as a mechanistic blueprint of reasoning itself.<n>We introduce three novel RL strategies that dynamically perform targeted credit assignment to critical nodes.
  arXiv  Detail & Related papers  (2025-10-15T13:49:51Z)
• Understanding the Effects of Domain Finetuning on LLMs [60.874016669351874]
  We present the first systematic study of domain-specific fine-tuning in large medical language models.<n>Our analysis reveals that fine-tuning modifies only a small subset of the representational subspace.<n>To interpret these changes in subspaces, we propose tuning vectors, which explicitly capture the directional parameter shifts induced by fine-tuning.
  arXiv  Detail & Related papers  (2025-10-10T13:14:06Z)
• Decomposing MLP Activations into Interpretable Features via Semi-Nonnegative Matrix Factorization [17.101290138120564]
  Current methods rely on dictionary learning with sparse autoencoders (SAEs)<n>Here, we tackle these limitations by directly decomposing activations with semi-nonnegative matrix factorization (SNMF)<n>Experiments on Llama 3.1, Gemma 2 and GPT-2 show that SNMF derived features outperform SAEs and a strong supervised baseline (difference-in-means) on causal steering.
  arXiv  Detail & Related papers  (2025-06-12T17:33:29Z)
• Understanding Task Vectors in In-Context Learning: Emergence, Functionality, and Limitations [19.539276425108987]
  This work proposes the Linear Combination Conjecture, positing that task vectors act as single in-context demonstrations formed through linear combinations of the original ones.<n>We show that task vectors naturally emerge in linear transformers trained on triplet-formatted prompts through loss landscape analysis.<n>We predict the failure of task vectors on representing high-rank mappings and confirm this on practical LLMs.
  arXiv  Detail & Related papers  (2025-06-10T17:59:31Z)
• Steering LLM Reasoning Through Bias-Only Adaptation [12.246105935814683]
  We show that training a single $d$-dimensional steering vector per layer with reinforcement learning, while freezing all base weights, matches the accuracy of fully RL-tuned reasoning models on mathematical-reasoning tasks.
  arXiv  Detail & Related papers  (2025-05-24T13:55:38Z)
• Steering Risk Preferences in Large Language Models by Aligning Behavioral and Neural Representations [4.029252551781513]
  We propose a principled approach for uncovering steering vectors.<n>We focus on extracting latent risk preferences from large language models.<n>We show that the resulting steering vectors successfully and reliably modulate LLM outputs in line with the targeted behavior.
  arXiv  Detail & Related papers  (2025-05-16T18:23:10Z)
• Illusion or Algorithm? Investigating Memorization, Emergence, and Symbolic Processing in In-Context Learning [50.53703102032562]
  Large-scale Transformer language models (LMs) trained solely on next-token prediction with web-scale data can solve a wide range of tasks.<n>The mechanism behind this capability, known as in-context learning (ICL), remains both controversial and poorly understood.
  arXiv  Detail & Related papers  (2025-05-16T08:50:42Z)
• SEAL: Steerable Reasoning Calibration of Large Language Models for Free [58.190800043449336]
  Large Language Models (LLMs) have demonstrated compelling capabilities for complex reasoning tasks via the extended chain-of-thought (CoT) reasoning mechanism.<n>Recent studies reveal substantial redundancy in the CoT reasoning traces, which negatively impacts model performance.<n>We introduce SEAL, a training-free approach that seamlessly calibrates the CoT process, improving accuracy while demonstrating significant efficiency gains.
  arXiv  Detail & Related papers  (2025-04-07T02:42:07Z)
• One-shot Optimized Steering Vectors Mediate Safety-relevant Behaviors in LLMs [21.2431937128876]
  We propose optimizing steering vectors through gradient descent on a single training example.<n>We find that the resulting SVs effectively mediate safety-relevant behaviors in multiple models.<n>We extend work on "emergent misalignment" and show that SVs optimized to induce a model to write vulnerable code cause the model to respond harmfully.
  arXiv  Detail & Related papers  (2025-02-26T06:13:01Z)
• Vector-ICL: In-context Learning with Continuous Vector Representations [75.96920867382859]
  Large language models (LLMs) have shown remarkable in-context learning capabilities on textual data.<n>We explore whether these capabilities can be extended to continuous vectors from diverse domains, obtained from black-box pretrained encoders.<n>In particular, we find that pretraining projectors with general language modeling objectives enables Vector-ICL.
  arXiv  Detail & Related papers  (2024-10-08T02:25:38Z)
• Activation Scaling for Steering and Interpreting Language Models [55.59689963561315]
  We argue that successfully intervening on a model is a prerequisite for interpreting its internal workings. We establish a three-term objective: a successful intervention should flip the correct with the wrong token and vice versa. Using gradient-based optimization, this objective lets us learn (and later evaluate) a specific kind of efficient and interpretable intervention.
  arXiv  Detail & Related papers  (2024-10-07T12:01:32Z)
• In-Context Convergence of Transformers [63.04956160537308]
  We study the learning dynamics of a one-layer transformer with softmax attention trained via gradient descent. For data with imbalanced features, we show that the learning dynamics take a stage-wise convergence process.
  arXiv  Detail & Related papers  (2023-10-08T17:55:33Z)
• Interpretability at Scale: Identifying Causal Mechanisms in Alpaca [62.65877150123775]
  We use Boundless DAS to efficiently search for interpretable causal structure in large language models while they follow instructions. Our findings mark a first step toward faithfully understanding the inner-workings of our ever-growing and most widely deployed language models.
  arXiv  Detail & Related papers  (2023-05-15T17:15:40Z)
• Extracting Latent Steering Vectors from Pretrained Language Models [14.77762401765532]
  We show that latent vectors can be extracted directly from language model decoders without fine-tuning. Experiments show that there exist steering vectors, which, when added to the hidden states of the language model, generate a target sentence nearly perfectly. We find that distances between steering vectors reflect sentence similarity when evaluated on a textual similarity benchmark.
  arXiv  Detail & Related papers  (2022-05-10T19:04:37Z)
• Layer-wise Analysis of a Self-supervised Speech Representation Model [26.727775920272205]
  Self-supervised learning approaches have been successful for pre-training speech representation models. Not much has been studied about the type or extent of information encoded in the pre-trained representations themselves.
  arXiv  Detail & Related papers  (2021-07-10T02:13:25Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.